$ grep -c "costs" run-report.md▊

"Cost if wrong" on every ruling

Every decision in our run reports carries one extra clause: what it costs if it turns out wrong. It takes ten seconds to write. It is the reason two architecture rules could be reversed in a sentence instead of an argument.

Andrii Siuta · September 2026 · ~4 min read decisions AI agents engineering practice

01. The format

When a multi-agent run is under way, the dispatcher makes dozens of small calls: run these tasks in parallel, accept this departure from the spec, ship this despite a missed target. Each is written into the run's ledger as it is made, and each ends with its price: 

rulings, as writtenRuling: Tasks 1-3 run in parallel in the one shared tree.
  Costs a transient verify failure on another task's half-edit; re-run.

Ruling: ship per-project Jest at 1.49x worst case against a 1.4x bar.
  Costs a shared-library PR about a minute; the average PR is 17-28 % faster.

Ruling: skip the remote cache, revisit after three months of affected numbers.
  Costs CI minutes on cold agents until then; nothing else depends on it.

One CI story alone carries 25 rulings this way. A single UI-kit wave carries 47, all waiting for a human to confirm.

02. Why it made reversals cheap

Twice during an Nx migration we reversed an architecture rule. The first put a domain's store in its feature library because nothing else read it. The second said that anything two domains share is lifted into shared. Both were reasonable, and both turned out wrong: the second left shared holding 442 of 1,631 library files, so one change there affected 99 % of test suites.

Neither reversal was an argument, because the cost had been written down when the rule was made. Nobody had to reconstruct why the rule existed, or defend it. The question was only whether the written price had come due, and it had.

A decision with its price attached is a decision you can afford to change.

03. What it does for agents

  • It survives context loss. A new session reads the ruling and its cost instead of re-deciding, or worse, deciding differently.
  • It makes the human review fast. Forty rulings with prices read in minutes; the expensive ones stand out.
  • It forces the dispatcher to think one step ahead, which is most of what a good ruling needs.

Write it in the same line as the decision, in the past tense of the failure: "costs a re-run", "costs a minute per shared PR", "costs two commits dropped at the PR". If you cannot name the cost, that is the ruling to ask a human about.